Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
Note
This article describes features used in agents or agent flows powered by the standard harness.
Tools are building blocks that let your agent interact with external systems. Tools expand what your agent can do, and your agent can perform various actions in response to user requests or autonomous triggers. Each tool represents a specific capability that your agent can perform. For example, you can equip your agent with tools that perform tasks like:
- Send emails by using the Office 365 Outlook connector
- Check the current weather conditions and forecasts
- Read and write data from Dataverse
- Read and post messages to Teams
Tools in an agent or capabilities in a plugin?
The tools you add in Copilot Studio belong to your agent. Your agent decides when to use them, and you manage them in Copilot Studio. A Microsoft 365 Copilot plugin is different: it packages skills, tools such as MCP servers, and agents so people can find, install, and manage them together in supported Copilot experiences.
| Questions and examples | Tools in your agent | A Microsoft 365 Copilot plugin |
|---|---|---|
| Use it when? | You're designing one agent experience and the tools only make sense inside it. | You want people to use a capability directly in the Copilot experiences they already work in. |
| Who decides when a tool runs? | Your agent's orchestration, topics, and settings. | Copilot, in each supported experience. |
| Where people use it? | The channels you publish your agent to, such as Teams. | Supported Copilot experiences. Each experience uses the parts of the plugin it supports. |
| Who manages it? | Makers in Copilot Studio. Power Platform data policies control which tools makers can add. | Admins, in the Microsoft 365 admin center. |
| Example | Contoso's IT help desk agent creates tickets and resets passwords in Teams. | Contoso's customer account plugin packages a skill that prepares customer briefs and an MCP server that looks up accounts in the company's CRM. Sellers ask Copilot to "prepare a brief for Fabrikam." |
Example: Tools in an agent (Contoso IT help desk agent)
Contoso IT builds an agent in Copilot Studio and publishes it to Teams. The agent uses a ServiceNow connector to look up and create tickets, an agent flow that resets a password after a manager approves, and a prompt that summarizes ticket history. The agent's topics and orchestration decide when each tool runs, and the password reset asks the user to confirm first. The IT team manages the agent in Copilot Studio, and Power Platform data policies control which connectors makers can use. Tools in an agent fit here because they only make sense inside this help desk conversation, which Contoso designs from start to finish.
Example: A plugin (Contoso customer account plugin)
Contoso Sales packages a plugin that contains a skill that prepares a customer brief from the team's template, an MCP server that looks up accounts in Contoso's CRM, and, optionally, a declarative agent for account reviews. Contoso publishes the plugin to its organization, and admins make it available to the sales team in the Microsoft 365 admin center. Sellers ask Copilot in supported experiences to "prepare a brief for Fabrikam." Copilot uses the skill and the CRM tool, and sellers don't open a separate agent. The first time, they connect their CRM account. A plugin fits here because the capability belongs in the Copilot experiences sellers already use, and admins manage it in one place.
To learn more about plugins, see Plugins.
Mechanisms for adding tools to agents
Note
If a tool you expect isn't available in the Add a tool pane, your organization's policies might restrict the tools available to you. Contact your administrator for help.
Your agent can use tools in two ways:
Generative orchestration (automatically): Generative orchestration is on by default. Your agent picks the best tool, topic, or knowledge source for each request.
Example: A user enters "My laptop won't connect to VPN," and the agent chooses the ServiceNow tool to open a ticket.
From a topic (explicitly): You call the tool from a node in a topic. If you turn off generative orchestration (classic mode), your agent responds only through topics, but topics can still call tools.
You can add these types of tools:
| Tool type | What it does |
|---|---|
| Prebuilt connector | Connects to one of thousands of Microsoft and non-Microsoft services through Power Platform connectors. |
| Custom connector | Connects to your own service through a Power Platform connector that you define. For your agent to use it, the connector must be viewable and shared across your organization. |
| Agent flow | Runs a flow of one or more actions. |
| Prompt | Runs a single-turn model prompt. It can use knowledge you provide and generate code to analyze data. |
| REST API | Calls the endpoints and methods you select from a REST API. |
| Model Context Protocol (MCP) | Connects to an MCP server and uses its tools and resources. |
| Computer use | Operates websites and desktop apps through their user interface: selecting buttons, choosing menus, and entering text. |
Two other options give your agent tool-like behavior:
Azure Bot Service skills make a set of related tools from an Azure Bot Service bot available to your agent. Client tools send an event activity to the client app, which performs an action and returns a response.
Note
Skills in agents powered by the standard harness are Azure Bot Service skills. They're different from agent skills, which you can add to agents powered by the GitHub Copilot harness.
Create and add a new tool at agent level
When you add a tool to your agent, your agent can use it automatically in any conversation.
In Copilot Studio, select Agents, and then select your agent.
Go to the agent's Tools page.
Select Add a tool.
Select the type of tool. To see every type, including REST API and Custom connector, select See all.
Note
If your organization blocks a type of tool, you can't select it. For more information, see Tools blocked by your organization's policies.
Configure the tool. The steps depend on the type. For example, for a Prompt tool, write the prompt instructions, select a model, and add input parameters by selecting + Add content. You can get help from the prompt assistant or start from a prompt template.
Select Save, and then select Add and configure.
The tool's configuration page opens, and the tool now appears on the agent's Tools page.
Note
If your organization blocked a tool due to data policies, you can't select it. Learn more in Tools blocked by your organization's policies.
Perform the configuration steps specific to the type of tool you selected. For example, if you select Prompt, perform the following steps:
- Define the prompt instructions
- Select a model
- Specify input parameters (by selecting + Add content)
You can also get help by typing in the Prompt assistant field.
As an alternative, use a predefined template to create a prompt by selecting prompt template.
Select Save to create the new tool.
Select Add and configure. The tool is added to your agent. The configuration page for your tool appears. View and make changes to your tool configuration here.
Your new tool appears on the Tools page for the agent.
Tools blocked by your organization's policies
Admins can use governance policies to limit which tools and connectors are available in Copilot Studio. When a data policy blocks a tool, it appears unavailable in the Add a tool pane. To see which policy blocks it, select the information icon next to the tool. If you need a blocked tool, contact your administrator.
Admins manage plugins, including the skills and MCP servers they contain, in the Microsoft 365 admin center. Power Platform data policies still control which tools makers can add in Copilot Studio.
Learn more in Configure data policies for agents.
View and make changes to your tool configuration
To view or change a tool's settings, go to your agent's Tools page and select the tool. The configuration page has three sections: Details, Inputs, and Completion.
MCP servers have a different configuration page. Instead of Inputs and Completion, it has Tools and Resources sections that list what the MCP server provides. For more information, see Connect your agent to an existing MCP server.
Details
| Setting | What to know |
|---|---|
| Name | Appears in the agent's list of tools. Use a name that says what the tool does. |
| Description | Generative orchestration uses the description to decide when to use the tool. Say what the tool does and when to use it. |
| Allow agent to decide dynamically when to use the tool | When selected, your agent can choose this tool on its own. When cleared, your agent uses the tool only when a topic calls it. Selected by default when generative orchestration is on. |
| Ask the end user before running | Asks the user to confirm before the tool runs. Set to No by default. |
| Authentication | Runs the tool with the user's credentials (End user, the default) or the maker's credentials (Maker-provided). |
| Authentication description | Optional. Tells users what they're being asked to sign in to. |
Example: Contoso IT describes the password reset flow as "Resets the user's password after their manager approves. Use when a user is locked out or asks to reset their password." It also sets Ask the end user before running to Yes.
Inputs
Inputs are the values your tool needs. Each input appears as a row in a table.
By default, Fill using is set to Dynamically fill with AI. Your agent tries to find each value in the conversation, such as the user's message. If it can't find the value, it asks the user.
To control how an input is collected, select Customize. You can set:
- Display name and Description: how the input appears to the user.
- Identify as: how to interpret the user's answer, for example as text or as a predefined entity.
- Retry logic: whether to ask again if the agent doesn't recognize the answer.
- Input validation: extra checks beyond the entity type's default.
To skip asking the user, set Fill using to Custom value, and then enter a value, select a variable, or write a Power Fx formula.
Note
After you refresh a connector action, its output type might stay unresolved until you enter all required inputs. Enter the values, save the topic, and then check the output variables again.
Completion
Choose what happens after the tool runs. Under After running, select one of these options:
- Don't respond (default): Your agent includes the tool's output in its own response.
- Write the response with generative AI: AI writes a response based on the tool's output.
- Send specific response: You write a response template and insert output variables. You can format it with Power Fx.
- Send an adaptive card: You send an interactive card with buttons and actions.
You can also choose which output variables your agent and other tools can use.
Tool selection and input collection
Your agent decides whether to use a tool based on:
- The tool's name and description.
- The conversation so far.
- What the user is asking for.
- The inputs and outputs the tool has.
- Tools it already used in the conversation.
Note
If a response must reproduce tool output, such as a link or card content, exactly, use Send specific response with variable insertion or Send an adaptive card. Use Don't respond only when the tool already sent the user-visible message. Write the response with generative AI can rephrase tool output and might not preserve links or card content exactly.
Configure which output variables to make available to the agent and other tools.
Your agent also uses the tool's description to write questions that collect missing inputs, so you don't need to add question nodes for each one.
By default, a tool returns its results to your agent, which then writes a response. You can instead have the tool respond to the user directly.
Tip
When your agent asks AI-generated questions, let users know that AI wrote parts of the conversation. For example, add a message to the Conversation Start system topic.
Call an existing tool from within a topic
Use a topic when you want a tool to run at a specific point in a conversation.
In Copilot Studio, go to your agent's Topics page.
Create a topic and name it, for example, Create a ticket.
Add trigger phrases, such as "open a ticket," "report a problem," and "I need IT help."
Select Add node (+), and then select Add a tool.
Select the tool. Tools are grouped on the Basic tools, Connector, and Tool tabs.
Select Save.
MCP connector information
For the MCP connector, you can view the names and descriptions of the MCP tools and resources that the MCP connector makes available. The information for the tools and resources appears in a table, with one line for each tool.
Authentication considerations for tools
Some tools, such as prompts or tools that call Dataverse, need authentication. Your agent runs tools in the user's context, and tools don't run unless authentication is turned on. Choose one of these methods:
- End user credentials: Users only get data they're allowed to see.
- Maker-provided credentials: The tool uses your credentials. Use this method for shared resources, or when users shouldn't need their own access.
Test your tool
By using generative orchestration, your agent uses the tool when it decides the tool fits the request. To test it, ask a question the tool should handle. In the following example, Contoso IT types "I'm locked out of my account" and checks that the agent asks for confirmation and then runs the password reset flow. You can also test a tool by calling it from a topic.
Turn a tool on or off in your agent
A new tool is on by default. Turning a tool off stops your agent from using it, but keeps it connected so you can turn it back on later.
Go to your agent's Tools page and select the tool.
At the top of the configuration page, use the Enabled toggle.
Select Save.
Delete a tool from your agent
Go to your agent's Tools page.
Hover over the tool, and then select the three dots (…).
Select Delete, and then select Delete again to confirm.
Tip
You can delete tools only from an agent's Tools page, not from the main Copilot Studio Tools page.
Limitations on tools in agents
- By using generative orchestration, an agent can have up to 128 tools. For the best results, use 25 to 30.
- In multi-agent orchestration, each child agent can have its own 128 tools.